Add hparams API

[JamieMair]

Continues work from #77.

EDIT: This now works! Closes #87

[oxinabox]

I am excited about this feature.

[JamieMair]

@oxinabox @nomadbl I have finished the feature. The way that dictionary was serialised was actually wrong in the protobuf code, so I made it match the output of tensorboardX now. Thankfully, this fixed the issue of only one parameter showing up.

I have added some docs as well on how to use it, along with an example script. Should be simple enough as users should only really interact with write_hparams!.

Note, I also bumped the version of the package as well. Should be no breaking changes as this only adds functionality.

[nomadbl]

There's some really great work here, kudos.
However it looks like this points to a bug in Protobuf.jl.
The functions which are edited here already exist in https://github.com/JuliaIO/ProtoBuf.jl/blob/master/src/codec/encode.jl
Please file an issue there, as your understanding of the inner workings there is better than mine.
We should probably wait until we resolve it there and then we wouldn't need to add functions on our end which may break in the future if any changes to Protobuf.jl are made.

[JamieMair]

There's some really great work here, kudos.

Thanks!

However it looks like this points to a bug in Protobuf.jl.

I have opened an issue (JuliaIO/ProtoBuf.jl#233) and PR (JuliaIO/ProtoBuf.jl#234), but these would be breaking changes for any downstream users, so I think that there is a chance that they might not accept the changes - it may take a while for this to go anywhere.

We should probably wait until we resolve it there and then we wouldn't need to add functions on our end which may break in the future if any changes to Protobuf.jl are made.

As the added methods have strict types, only applying to the HParams plugin, how would you feel about merging these changes now, as any changes in ProtoBuf may take a while/not happen? I have added a unit test with what the exact serialisation and deserialisation should be, which will make it easier to transition to any upstream fixes while verifying that tensorboard can still parse the output.

[nomadbl]

I think with the precautions you have added it is reasonable to merge this now.
Thanks for taking care of the Protobuf issue and PR as well!

I have added another suggestion and approved the PR. Merge as you see fit.

[JamieMair]

@nomadbl Thanks for the quick review and guidance. I'll updated from master and will merge as soon as the unit tests pass.

[JamieMair]

@nomadbl I have realised that I don't actually have write access, I think someone else would need to merge the changes. Thanks :)